1 Ecology shapes epistasis in a genotype-phenotype-fitness map for stick insect color Patrik Nosil 1,2 , Romain Villoutreix 1 , Clarissa F. de Carvalho 1 , Jeffrey L. Feder 3 , Thomas L. Parchman 4 , and Zach Gompert 2 1 Centre d’Ecologie Fonctionelle et Evolutive, Centre National de la Recherche Scientifique, Montpellier, 34293, France 2 Department of Biology, Utah State University, Utah 84322, USA 3 Department of Biological Sciences, University of Notre Dame, Notre Dame, Indiana 46556, USA 4 Department of Biology, University of Nevada Reno, Nevada, 89557, USA Lead contact: [email protected]Abstract Genetic interactions such as epistasis are widespread in nature and can shape evolutionary dynam- ics. Epistasis occurs due to non-linearity in biological systems, which can arise via cellular processes that convert genotype to phenotype and via selective processes that connect phenotype to fitness. Few studies in nature have connected genotype to phenotype to fitness for multiple potentially in- teracting genetic variants. Thus, the causes of epistasis in the wild remain poorly understood. Here, we show that epistasis for fitness is an emergent and predictable property of non-linear selective processes. We do so by measuring the genetic basis of cryptic colouration and survival in a field ex- periment with stick insects. We find that colouration exhibits a largely additive genetic basis, but with some effects of epistasis that enhance differentiation between colour morphs. In terms of fit- ness, different combinations of loci affecting colouration confer high survival in one host-plant treatment. Specifically, non-linear correlational selection for specific combinations of colour traits in this treatment drives the emergence of pairwise and higher-order epistasis for fitness at loci un- derlying colour. In turn, this results in a rugged fitness landscape for genotypes. In contrast, fitness epistasis was dampened in another treatment, where selection was weaker. Patterns of epistasis that are shaped by ecologically based selection could be common, and central to understanding fitness landscapes, the dynamics of evolution, and potentially other complex systems. Genes that control adaptive traits have now been identified in many organisms 1,2 and some pioneering work has even connected genotype to phenotype to fitness (i.e., a genotype-phenotype-fitness map) for in- dividual genes 3–8 . However, adaptation may often involve multiple genes 9–11 , with potential interactions
30
Embed
Ecology shapes epistasis in a genotype-phenotype-fitness ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Ecology shapes epistasis in a genotype-phenotype-fitness map for stick insect color
Patrik Nosil1,2, Romain Villoutreix1, Clarissa F. de Carvalho1, Jeffrey L. Feder3, Thomas L. Parchman4,
and Zach Gompert2
1Centre d’Ecologie Fonctionelle et Evolutive, Centre National de la Recherche Scientifique, Montpellier,
34293, France2Department of Biology, Utah State University, Utah 84322, USA3Department of Biological Sciences, University of Notre Dame, Notre Dame, Indiana 46556, USA4Department of Biology, University of Nevada Reno, Nevada, 89557, USA
Genetic interactions such as epistasis are widespread in nature and can shape evolutionary dynam-
ics. Epistasis occurs due to non-linearity in biological systems, which can arise via cellular processes
that convert genotype to phenotype and via selective processes that connect phenotype to fitness.
Few studies in nature have connected genotype to phenotype to fitness for multiple potentially in-
teracting genetic variants. Thus, the causes of epistasis in the wild remain poorly understood. Here,
we show that epistasis for fitness is an emergent and predictable property of non-linear selective
processes. We do so by measuring the genetic basis of cryptic colouration and survival in a field ex-
periment with stick insects. We find that colouration exhibits a largely additive genetic basis, but
with some effects of epistasis that enhance differentiation between colour morphs. In terms of fit-
ness, different combinations of loci affecting colouration confer high survival in one host-plant
treatment. Specifically, non-linear correlational selection for specific combinations of colour traits
in this treatment drives the emergence of pairwise and higher-order epistasis for fitness at loci un-
derlying colour. In turn, this results in a rugged fitness landscape for genotypes. In contrast, fitness
epistasis was dampened in another treatment, where selection was weaker. Patterns of epistasis that
are shaped by ecologically based selection could be common, and central to understanding fitness
landscapes, the dynamics of evolution, and potentially other complex systems.
Genes that control adaptive traits have now been identified in many organisms 1,2 and some pioneering
work has even connected genotype to phenotype to fitness (i.e., a genotype-phenotype-fitness map) for in-
dividual genes 3–8. However, adaptation may often involve multiple genes 9–11, with potential interactions
2
among them. Here, we focus on epistasis, defined as interactions between genes, where the effects of an
allele at a locus depend on an allele (or alleles) at one or more other loci in the genome 12. Such epistasis
can make it difficult to predict evolution based on information from single genes alone 9,13,14, and has im-
plications for adaptation 11,12,15 and speciation 16–18. For example, epistasis can affect the evolution of com-
plex traits, sex and recombination 19, parasite and antibiotic resistance 20,21, reproductive isolation 16,17, and
missing heritability in human disease 9. Epistasis is fundamental for understanding the structure of fitness
landscapes 18,22,23, including their ruggedness and the number of adaptive peaks they contain, features that
shape evolutionary dynamics.
Epistasis can arise at two fundamental levels of biological organisation (Figure 1)10,24,25. First, cellular and
molecular processes can result in non-linearity in the conversion of genotypic to phenotypic variation 12,15,24, for example due to protein interactions and the complexity of metabolic and developmental net-
works 24 (a ‘non-linear genotype-phenotype map’ hypothesis, Fig. 1). Second, non-linear forms of pheno-
typic selection can cause epistasis for fitness at genes underlying trait variation 11,25,26 (a ‘fitness epistasis’
hypothesis). Critically, this can occur even if alleles contribute additively to trait variation, because when
it comes to fitness per se the effect of an allele can still depend on the genetic background in which it oc-
curs 25. For example, under stabilizing or disruptive phenotypic selection (i.e., common forms of non-lin-
ear selection) 27, the fitness effect of a mutation that additively increases a trait value (e.g., body size) will
depend on whether the mutation occurs in a genetic background where it moves the multi-locus genotype
closer or further from the selective optimum. Likewise, correlational selection for combinations of trait
values results in some underlying gene combinations having higher fitness than others, i.e., fitness epista-
sis rather than an additive relation between genotype and fitness 28. Parsing these two main causes of epis-
tasis is important because it determines whether interactions arise via inherent cellular features or through
variation in ecological factors.
Empirically, epistasis is difficult to study due to the vastness of genotype space and the challenge of con-
necting genotype to phenotype to fitness for multiple genetic variants. For example, in terms of genotype
space, with just five mutational steps separating two DNA sequences, there are 5! = 120 possible muta-
tional paths between them 21. Nonetheless, some progress has been made. A number of innovative combi-
natorial studies have engineered mutational steps that separate forms of single proteins and tested their fit-
ness effects 29. Some of these studies provide evidence for sign epistasis (i.e., alleles beneficial in one ge-
netic background are deleterious in another) and rugged fitness landscapes 21,29, as do experimental evolu-
tion studies in microbes 13,30. Most of this work has focused on simple pairwise genetic interactions such
that higher-order interactions remain poorly understood, despite the potential importance of the dimen-
3
sionality of the fitness landscape for evolutionary dynamics 18,22. In terms of genotype-phenotype-fitness
maps, a general understanding of how genetic variation is converted into phenotypes is accumulating
rapidly 1–9. In contrast, much less is known about the genetic basis of organismal fitness in the wild 1,3–8,
and very few studies have connected genotype to phenotype to fitness for combinations of loci (but see 11). Until more such studies emerge, the prevalence, causes, and predictability of genetic interactions will
remain unclear.
Connecting genotype to phenotype to fitness for multiple loci is not a trivial task. Moreover, in many sys-
prevents the existence of a range of genetic combinations whose fitness can be assayed 19,31. Thus, a key to
being able to test hypotheses concerning epistasis is the existence of recombination among adaptive ge-
netic variants, and the ability to measure fitness of different gene combinations. Here, we use segregating
genetic variation created by recombination to experimentally connect genotype to phenotype to fitness for
loci controlling cryptic colouration in an insect. This enables us to elucidate the causes of epistasis and to
infer features of the fitness landscape. In turn, our results inform why gene combinations are often pack-
aged into distinct units of biodiversity via suppressed recombination.
We study wingless, herbivorous Timema stick insects, which rely on crypsis for protection against visual
predators such as birds while resting on their host plants 32–36. Timema body colouration has thus evolved
to approximate the colours of the leaves and stems of their hosts (i.e., green versus brown / ‘melanistic’
morphs), and colour is a major axis of selection and adaptation in these insects 32,34,35 (Figures 2, 3). A re-
cent study revealed that colour in T. chumash is controlled by a moderate number (~5-7) of linked but re-
combining genetic variants that reside in a ~1 mega-base region of linkage group eight (LG8 hereafter) 37.
Accordingly, although T. chumash exhibits statistically distinguishable green and melanistic morphs, it
exhibits wide ranging colour variation overall, including individuals that are shades of yellow, pink, tan,
beige, and blue 37. Here, we leverage this segregating genetic variation for a range of colouration to test
for selection on combinations of colouration loci (i.e., fitness epistasis). We integrate our findings with
the fact that other Timema species exhibit more distinct colour morphs due to suppressed recombination
among colour loci 37. For example, morphs in T. cristinae are distinguished by a large (~10 mega-base)
region of suppressed recombination on LG8 named the ‘Mel-Stripe’ locus 35,37,38. We thus focus here ex-
clusively on the Mel-Stripe region (note that the majority of colour loci in T. chumash map to a ~1 mega-
base subset of the Mel-Stripe region) 37.
Results and Discussion
4
Genotype-phenotype map for cryptic colouration. We first tested the hypothesis of a non-linear geno-
type-phenotype map for colouration in T. chumash (see Methods for details). We did so using standard-
ized photos of 437 T. chumash for which we quantified body colour using red, green, and blue (R, G, B,
respectively) pixel values. Following the approach of Endler 39, we calculated chromatic contrasts as the
relative difference between: (1) red and green channels (RG = (R – G)/(R +G), a trait referred to as ‘RG’
hereafter) and (2) between green and blue channels (GB = (G – B)/(G + B), a trait referred to as ‘GB’
hereafter; Figure 2). This approach is based on the two most common differences between photoreceptor
signals, resulting from the wiring of visual systems 40,41. Such contrasts thus yield more biologically
meaningful results for comparing colour patterns than do raw RGB values 39. In addition, our methods
capture the major axes of variation in T. chumash colour space, given this species does not reflect ultravi-
olet spectra (see Supplementary Information and Supplementary Figure 1 for results using reflectance
data from T. chumash, modelled to avian photoreceptor sensitivities). These same photographed individu-
als were subsequently used in the mark-recapture experiment described below to estimate phenotypic se-
lection and the genetic basis of survival. Thus, from each individual we took a tissue sample before re-
lease into the field that allowed us to collect genotyping-by-sequencing (GBS) data for all individuals
(both those recaptured and not). Past work comparing individuals from whom a tissue sample was taken
to unperturbed individuals has shown that tissue sampling does not affect survival in the laboratory or
field 42.
The GBS data were first used for genetic mapping of colour (Figure 3). To this effect, we employed a
purely additive Bayesian multi-locus genome-wide association (GWA) mapping approach in GEMMA
that accounts for linkage disequilibrium (LD) among SNPs 43. This revealed that ~80% of the variation in
colour was explained by additive genetic variation, with narrow credible intervals on these estimates (Fig-
ure 3, Supplementary Figure 2). Consistent with past work, the SNPs associated with colour were concen-
trated in the Mel-Stripe locus region of the T. cristinae genome 37. We return to these individual SNPs in
more detail below when connecting genotype to fitness.
Given that most (i.e., ~80%) of the variation in colour was accounted for using an additive genetic model,
epistasis could at most explain 20% of colour variation. Past work on colouration in T. chumash using a
population from a different geographic site failed to detect evidence for epistasis 37. However, explicit
tests for marginal epistasis in the current data set using MAPIT 44 revealed statistical evidence of epistasis
for five SNPs in the Mel-Stripe region (Figure 3).
Consequently, we re-ran GEMMA analyses, this time including in the model the 10 possible pairwise in-
5
teractions between these five SNPs (i.e., including epistatic interactions). This revealed a noticeable in-
crease (~10%) in the percent variance explained (PVE) for RG, but not for GB where PVE was compara-
ble to that in the additive model (Figure 3). Moreover, the interaction between two pairs of SNPs was
consistently retained as colour-associated across Markov chain Monte Carlo (MCMC) model steps (pos-
terior inclusion probability ~1.0 for one SNP pair affecting RG and ~0.6 for one SNP pair affecting GB).
We also quantified genomic estimated breeding values (GEBVs; these quantify the total effect of genetic
markers on phenotype) for GEMMA models with and without interactions. The GEBVs from the two
analyses were highly correlated such that similar estimates emerge whether epistasis was included or not
(RG, r = 0.95, P < 0.001 GB, r = 0.98, P < 0.001; both tests two-tailed). However, visual inspection of the
results revealed that allowing for epistasis does slightly alter GEBVs in a manner that enhances differenti-
ation between green versus darkly coloured individuals (Figure 3). Finally, genomic prediction of colour
based only on genotype revealed modestly increased predictive power in models that included epistatic
interactions. In sum, colour exhibits a largely additive basis, but with some moderate additional effects of
epistasis.
Phenotype-fitness map for survival. Having established that colour has a largely additive basis, we de-
signed a mark-transplant-and-sequence experiment to estimate phenotypic selection on colour and to map
the genetic basis of survival, a core component of fitness (Figure 4). Our design reflects patterns of
Timema evolution and host-plant use in Southern California, where T. chumash occurs. These considera-
tions focus on two closely related species, T. chumash and T. podura, which are broadly and locally co-
occurring (i.e., sympatric). T. podura exhibits highly distinct green and melanistic morphs and its core
hosts, Ceanothus (C) and Adenostoma (A), select for green versus melanistic colouration, respectively 34,45. These hosts are also used, albeit more secondarily, by T. chumash. The combination of these hosts
(AC hereafter) is predicted to generate strong correlational selection for green versus melanistic coloura-
tion (i.e., high GB and low RG values, or low GB and high RG values, respectively), and against other
colours. In contrast, a core host of T. chumash is Cercocarpus, i.e., mountain mahogany (MM), which ex-
hibits wide colour variation, as do T. chumash individuals found on it 37. We thus collected T. chumash
from MM and transplanted them within the same area to adjacent, touching individuals of A and C (AC
treatment) and to MM (control). Notably, such a scenario is biologically realistic given that these three
hosts co-occur at small spatial scales throughout Southern California.
We applied individual and block-specific pen-marks on the abdomen of each T. chumash. On June 18th,
these individuals were released onto bushes representing the two treatments, in three paired blocks. On
June 21st, the survivors of the experiment were recaptured. Several past experiments have revealed that
6
this procedure results in minimal dispersal, with individuals that are not recaptured suffering mortality
and those recaptured representing survivors 36,42,46. We estimated selection coefficients and standardized
selection gradients on colour by comparing survivors to non-survivors (Supplemental Table 1 for details
including sample sizes for numbers recaptured).
We did not detect strong evidence for selection on MM (Figure 4, Supplementary Figure 3). In contrast,
our prediction of correlational selection on AC was supported. Specifically, the combination of A and C
generated correlational selection for either: (1) high GB and low RG values (i.e., green morphs), or (2)
low GB and high RG values (i.e., melanistic morphs). Consequently, there was strong selection against
individuals with colouration intermediate or otherwise mismatched from that mentioned above. This was
the case for the experiment overall, and within two of three blocks individually. Standardized linear, qua-
dratic, and correlational selection gradients on AC were generally in the range of 0.05 – 0.15 (Supplemen-
tal Table 2). Thus, selection was moderately strong, but within the range documented for other systems
and traits 27. Correlational selection such as documented here can result in some gene combinations hav-
ing higher fitness than others (i.e., fitness epistasis). We thus conducted further analyses that integrate the
results from genotype, phenotype, and fitness to test for the effect of interactions between SNPs on sur-
vival probabilities.
Integrating components of the phenotype-genotype-fitness map. Our next goal was to connect geno-
type to phenotype to fitness for the specific genetic regions (i.e., SNPs) associated with colour. The first
step in doing so was to return to the results from the analyses reported above for mapping colour, this
time focusing on the individual SNPs most strongly associated with colour. Specifically, we used the
aforementioned results from the additive GEMMA model to: (1) quantify the weight of evidence that
each individual SNP was associated with colour, and (2) estimate the number of genetic variants (i.e.,
quantitative trait nucleotides, QTN) controlling each colour trait 43. This was done by considering how of-
ten SNPs were retained as trait-associated across different MCMC steps in the GWA. The proportion of
such steps is termed the posterior inclusion probability, PIP hereafter, and reflects the weight of evidence
that a SNP is associated with colouration. In the case of multi-genic control with recombination among
loci, the one or few SNPs that best tag each causal variant are expected to consistently be trait-associated
across MCMC steps (i.e., exhibit high PIP values). In turn, PIP values across SNPs sum to the number of
total causal variants (i.e., even if the causal variants are not unambiguously identified, the number of such
variants can be estimated) 43.
These analyses revealed that ~5 genetic variants control RG (posterior mean and s.d. = 4.63 +- 0.81), and
7
~4 control GB (posterior mean and s.d. = 3.94 +- 1.03). Overall, we estimated that ~6-7 genetic variants
control colouration (posterior mean and s.d. = 6.51 +- 1.15), because some, but not all, colour-associated
SNPs affected both traits. Notably, LD among the top colour-associated SNPs was generally modest, in-
dicative of recombination between them (Supplementary Figure 4). Given these results, we focused our
fitness analyses on the five most strongly colour-associated SNPs, with PIPs > 0.70 and minor allele fre-
quencies (MAF) greater than 0.05. Notably, these five SNPs were 478-672 times more likely to be colour-
associated than to have no effect on colour (Supplemental Table 3 for statistics). Finally, generation of a
new de novo chromosome-level genome assembly for T. chumash confirmed that these SNPs are in syn-
teny between T. cristinae and T. chumash, and exist as a single copy in the T. chumash genome (Supple-
mentary Figure 10).
We used Bayesian multiple regression with variable selection and model averaging to connect genotype
to phenotype to fitness. Our dependent variable was survival probability (i.e., expected fitness) inferred
from the analyses of phenotypic selection (i.e., each individual was assigned a survival probability based
on its colour score and the selection analyses). Our independent variables were the main (i.e., additive) ef-
fects of the five SNPs and their possible interactions (i.e., epistasis). In this context, two-way interactions
represent pairwise epistasis and other interactions represent higher-order epistasis.
The full results are depicted in Figure 5. In the AC treatment, we found additive and epistatic effects on
fitness, with the latter involving marked two- and three-way epistatic interactions. Thus, we estimated that
the number of effects of epistasis on survival on AC was ~6, with ~2-3 stemming from pairwise interac-
tions and ~2-3 stemming from three-way interactions. Notably, predictive power in the AC treatment
from leave one-out cross-validation was ~64% better for a model with both additive and epistatic effects
than for a model with only additive effects (both effects, predictive r = 0.60, 95% CI = 0.50-0.68, r2 =
0.36, P < 0.001, additive effects only, predictive r = 0.47, 95% CI = 0.35-0.57, r2 = 0.22, P < 0.001; both
tests two-sided; Table 1).
To test if these results could arise from chance (i.e., from no true association between genotype and fit-
ness), we ran null simulations that repeated the analyses 100 times using five randomly drawn SNPs that
were matched for MAF with the colour-associated SNPs. The null model simulations revealed that our re-
sults from AC were unlikely to arise by chance (P < 0.01 for main effects and two- and three-way interac-
tions; one-sided test). In contrast, additive and epistatic effects estimated in the MM treatment could be
explained by chance (P > 0.05 for all model terms; one-sided test), consistent with the weaker phenotypic
selection on MM than AC. Finally, these results were robust to other methods of analysis such as
8
Bayesian ridge and lasso regression, and could not be explained by dominance within SNPs (Supplemen-
tary Figure 5-8). We note that the genetic architecture of colour itself does not differ between treatments.
Consequently, our results are consistent with non-linear selection (specifically correlational selection)
driving the emergence of fitness epistasis on AC, and an absence of appreciable selection and fitness epis-
tasis on MM. Thus, fitness epistasis was predictable based on patterns of ecologically based selection.
Moreover, these effects of epistasis can be understood at the level of underlying pairs and triplets of
SNPs. One example in the AC treatment is shown in Figure 5, where alleles at two interacting SNPs have
been coded by whether they cause colouration to become green (G) or more melanistic (M). It can be seen
that high fitness is associated with having multiple copies of either G or M, with low fitness of genotypes
that combine these alleles (e.g., to result in intermediate colouration). Such effects can be extended be-
yond pairs of SNPs to higher-order interactions. For example, Supplementary Figure 9 uses a three-way
fitness interaction to illustrate how the survival effects of two SNPs that affect only GB depend on a third
SNP that affects both GB and RG. In this case, the effects of the first two SNPs on lowering GB scores
only improve survival if they are found in a genetic background at the third SNP that increases the RG
score to result in more melanistic colouration. Thus, epistasis for fitness can be understood predictably via
observed patterns of selection on colour in the transplant experiment.
Inference of fitness landscapes and their ruggedness. We next connected the detected effects of epista-
sis to the structure of the adaptive landscape. Specifically, we inferred the ruggedness of the fitness land-
scape (where peaks denote high fitness and valleys represent regions of low fitness) in our transplant ex-
periment using the unique fitness expectation for each genotype provided by our Bayesian regression
model. The multi-dimensional fitness landscape in each experimental treatment is depicted in Figure 5,
where nodes denote genotypes, node size represents sample size (where the smallest nodes represent
genotypes that were not observed in our sample), edges/lines connect genotypes that differ by one substi-
tution, and colours denote relative fitness. Visual inspection of the landscape suggests greater ruggedness
on AC than MM, consistent with the observed greater fitness epistasis on AC. Analyses executing random
walks on the landscape from different starting points confirm that this visual intuition is correct, with
greater ruggedness metrics on AC (Figure 5).
Supporting analyses and pleiotropic effects of colour-associated loci. Thus far, our fitness analyses
have focused on colour-associated SNPs, and we have implicitly assumed that the main phenotypic effect
of these SNPs is on colour, not other traits. In the supplementary materials and Table 1, we report addi-
tional analyses that explore and relax these assumptions, including a test for pleiotropic effects of colour-
9
associated SNPs on other (unmeasured) traits affecting survival (analogous to that introduced by Renni-
son and colleagues and applied to stickleback fish) 47. These results show that our finding of fitness epis-
tasis appears robust to different analytical approaches, but stronger results are obtained when considering
selection acting through colour phenotype than when considering genotype alone. They also maintain
clear evidence for selection on colour phenotype, but also suggest some pleiotropic effects of colour loci
on other (unmeasured) traits influencing fitness.
Conclusions. Epistasis can arise via inherent cellular features or through variation in ecological factors.
In our experiment, epistasis for fitness arose predictably as an emergent property of ecological variation
in natural selection. This result informs three core issues in biology: (1) the predictability and repeatabil-
ity of evolution, (2) exploration on adaptive landscapes, and (3) the packaging of multi-locus genetic vari-
ation into distinct units of biological diversity.
First, debate exists about the role of epistasis in the predictability and repeatability of evolution. Studies
of proteins suggest that if evolution were repeated from the same starting point (i.e., genetic background),
epistasis might increase the predictability of the mutational path taken to a given endpoint 21,29. Specifi-
cally, deleterious gene combinations constrain the number of paths that are accessible to selection, caus-
ing predictable evolution. However, evolution will often proceed from different starting points, for exam-
ple in variable genetic backgrounds of different populations and species. In such cases, genes with strong
epistatic effects may only function well in a narrow range of genetic backgrounds, reducing their repeated
use 12. Our results add an important component to this debate. A major cause of epistasis in our study was
selection itself. Thus, a key determinant of our ability to predict evolution might be our understanding of
selection and its ecological causes, rather than only the cellular features that create epistasis in the geno-
type to phenotype map.
Second, our results shed light on the exploration of fitness landscapes. Specifically, how do populations
traverse fitness valley to find global fitness peaks and avoid getting stuck on local optima 22,25? A famous
solution offered by Wright’s ‘Shifting Balance Theory’ 23,48 invokes a delicate balance of migration, drift,
and inter-demic selection, which may be difficult to achieve 49 (but see 50). A potentially more general so-
lution involves the stability of the landscape. If the environment is not static but rather fluctuates, then a
peak at one point in time can become a valley at another, and vice-versa. Thus, valleys are temporary and
crossable at certain points in time, i.e., the landscape is more a shifting ‘seascape’ 26. The weaker selection
on MM in our experiment could maintain standing genetic variation and help bridge peaks offered by AC,
facilitating exploration of the fitness landscape. Other forms of fluctuating selection, such as negative fre-
10
quency-dependent selection, could further enable the exploration process 26, and indeed such selection has
been documented in Timema 36. Thus, the ecologically mediated epistasis documented here may enable
the exploration of fitness landscapes in Timema. Given the ecological complexity of nature in time and
space 14,51, similar processes likely apply to other organisms.
Third, our results increase understanding of the processes that package genetic variation into distinct units
of biodiversity, such as morph or species. Divergence into such units is facilitated by reduced recombina-
tion 19,38 and by reproductive incompatibility 16,52. Epistasis may play a role in both 16,17. The alternative
gene combinations favoured here could generate a seed of linkage disequilibrium that promotes the evo-
lution of further reduced recombination between colour genes, and enhances the efficacy of selection for
such reduced recombination. Indeed, this may have occurred in T. cristinae, a relative of T. chumash that
feeds primarily on the hosts AC and exhibits reduced recombination between colour genes 36,38. Moreover,
our genome comparison here revealed that a chromosomal inversion on LG8 distinguishes these two
species (Supplementary Figure 10). Finally, we note that specific combinations of traits and genes were
selected against. Thus, natural selection itself may create ecological incompatibilities between forms that
are analogous to the classical Dobzhansky-Muller genetic incompatibilities that cause hybrid dysfunction 17. Although our focus was on morphs, similar processes could apply to ecotypes or species, as demon-
strated under semi-natural conditions in stickleback fish 11. Whether distinct units of diversity form could
be mediated by the temporal and spatial scales at which selection (and thus fitness epistasis) fluctuates;
fluctuations promote peak shifts but counteract consistent pressure for divergence.
Although experimental studies in nature akin to ours are few, genetic interactions have been studied ex-
tensively in several other contexts. For example, epistasis has been reported in experimental evolution
studies in the lab 30,53, work on protein evolution 12,21, and in genome-wide scans of population genetic pat-
terns 54,55. Ideally, future work would combine these approaches to parse the causes of epistasis across dif-
ferent contexts, thus testing the generality of the patterns reported here. Nonetheless, given the prevalence
of non-linear selection in nature 27, we expect fitness epistasis to be common. Thus, a collective body of
emerging evidence suggests that genetic interactions may be central to understanding biological diversifi-
cation, rather than being only complex second-order effects 56. Even more generally,
interacting components and non-linear functions are aspects of many biological, chemical, physical, and
social systems 57,58. Thus, evolutionary principles learnt from the study of genetic interactions may aid un-
derstanding of other complex systems.
ACKNOWLEDGEMENTS
11
We thank T. Reimchen, M. Joron, Luis-Miguel Chevin, and D. Ayala for discussion and comments on
previous versions of the manuscript, M. Muschick for help with photography and reflectance measure-
ments, and T. Oakley for lab space. The support and resources from the Center for High Performance
Computing at the University of Utah are gratefully acknowledged, as well as access to the High Perfor-
mance Computing Facilities, particularly to the Iceberg and ShARC HPC clusters, from the Corporate In-
formation and Computing Services at the University of Sheffield. The work was funded by a grant from
the European Research Council (EE-Dynamics 770826, https://erc.europa.eu/) and a grant from the Na-
tional Science Foundation of the USA (NSF DEB 1638768). The funders had no role in study design, data
collection and analysis, decision to publish, or preparation of the manuscript.
Author contributions. All authors (PN, RV, CFC, JLF, TLP, and ZG) conceived the project. RV, CFC,
TLP, and PN collected data. ZG led data analysis, aided by all authors. All authors (PN, RV, CFC, JLF,
TLP, and ZG) contributed to writing.
Data Availability Statement. DNA sequences have been deposited in the NCBI SRA (BioProject PR-
JNA656892). Other data, including colour measurements and results from the experiment,have been
archived in Dryad Digital Repository (doi:10.5061/dryad.2z34tmpjr). Correspondence and requests for
79. Paten, B. et al. Cactus: Algorithms for genome multiple sequence alignment. Genome Res. 21,
1512–1528 (2011).
80. Hickey, G., Paten, B., Earl, D., Zerbino, D. & Haussler, D. HAL: A hierarchical format for storing
and analyzing multiple genome alignments. Bioinformatics 29, 1341–1342 (2013).
Figure 1. Schematic of the hypotheses examined here. A) Epistasis can arise at two fundamental
levels; due to non-linear cellular processes in the genotype to phenotype map and due to non-lin-
ear selection in the phenotype to fitness map.
Figure 2. Objectively quantifying colour variation from digital photographs. A) Spectral sensitivities of
cones of a hypothetical tetrachromatic receiver (e.g., a bird 59). The human visible spectrum represents
cones that capture long wavelengths (L, red), medium wavelengths (M, green) and short wavelengths (S,
blue). Differences between L-M and M-S activities are the most common responses in colour perception 40,41. Relative differences between values of red and green (RG=(R – G)/(R +G)) and of green and blue
(GB=(G – B)/(G + B)) 39 extracted from digital photographs can be used as an approximation to this phys-
iological response. This approach was used considering T. chumash does not reflect UV (see Supplemen-
tary Information and Supplementary Figure 1). B) RG and GB are orthogonal measures and together cap-
ture the range of colour variation observed in T. chumash.
Figure 3. Genetics of cryptic colouration in T. chumash. A) Hyper-parameter estimates from GEMMA
models without versus with interactions between pairs of the five SNPs shown to have epistatic effects on
colouration in MAPIT (additive versus epistasis models hereafter). PVE = percent variance explained.
Vertical bars denote 95% credible intervals. The darker shading within each bar represents PGE = percent
of the PVE explained by SNPs with individually detectable effects. RG = red-green. GB = green-blue. B)
Posterior inclusion probabilities (PIPs) for SNPs for RG and GB colour traits. The grey, shaded region
represents the Mel-Stripe locus. C) Results from MAPIT analyses testing for epistatic effects on coloura-